Stay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
summary
Claims that CWE-1039 (malicious ML training) can be detected by automated model source-code analysis and stopped with both detection and hardening measures. Demonstrations use ONNX/OpenVINO-compiled edge models, adversarial generators (FGSM, BIM, Carlini–Wagner) and a denoising pipeline that restores a perturbed stop sign to correct classification. Proposed defenses include adversarial training, defensive distillation, watermarking/blockchain for training data, model compression, continuous monitoring and explainability tools like LIME/SHAP; expect a robustness-versus-accuracy tradeoff and active real-world incidents (Tesla sticker, Waymo cones).
AI-written summary · about
chapters (11)
- 0:00Introduction and scope
- 0:31Speaker background
- 2:13Understanding CWE-1039
- 5:27Regulatory and compliance
- 10:13Trusted AI in OT/ICS
- 13:36AI attack vectors
- 20:20OT/ICS risk scenarios
- 25:32Defense and mitigation
- 30:31Securing ML lifecycles
- 35:55Key takeaways
- 40:59Action plan and Q&A
related talks
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
36mAdversarial Machine Learning: How Secure Is ML?
28mSecurity of Machine Learning
44mComputers are Stupid: Protecting "AI" from Itself
46mA Short Introduction to Hacking Artificial Intelligence
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
31mSecurity Considerations for Services Using AI Models
01h 01mSecure AI: Integrating OWASP Principles in Machine Learning Model Development